Taizhen Cheung — Software Developer // PLAYER

Taizhen Cheung

ONLINE // v.2026

ABOUT ME PLAYER STATS

Who Am I?

I'm a Computer Science and Mathematics student at Hunter College (CUNY), graduating May 2027. My passion lives at the crossroads of artificial intelligence and real-world impact — from evaluating LLM fairness and benchmarking cutting-edge models, to engineering full-stack applications and AI emotional-companion platforms.

I'm driven by the challenge of making AI systems more reliable, unbiased, and accessible. Whether it's dissecting the reasoning patterns of GPT-4o and DeepSeek, architecting "FacetsRAG" retrieval pipelines, or building React apps that consume real-time APIs — I love turning complex problems into elegant, working software.

CS MAJOR

MATH MAJOR

2027 GRAD YEAR

Hunter College, CUNY

CURRENT MAIN SF6 // MASTER RANK

SAGAT

Zoner // The King of Muay Thai

MASTERY 65% • PLATFORM PC • CFN PLAYER_ONE

Outside of code I main Sagat in Street Fighter 6 — grinding ranked, labbing frame data, and chasing that next rank-up. Try typing "hadouken" or "tiger" anywhere on this page.

EXPERIENCE LOG 4 MISSIONS LOGGED

AI Trainer – Software Engineering & Coding Domain

Handshake – Remote

Jun 2025 – Present

Evaluated and scored responses from frontier large language models on multi-turn software-engineering and coding tasks, assessing correctness, helpfulness, and communication quality against detailed rubrics.
Authored written rationales for each score and conducted comparative preference evaluations across multiple candidate responses to identical coding prompts.
Maintained high inter-rater agreement through a rigorous calibration screening, generating high-quality human-feedback data to improve the reasoning and response quality of AI coding assistants.

AI Trainer – Image Processing Domain

Handshake – Remote

Nov 2025 – May 2026

Utilized Python-based tooling to evaluate and validate training datasets for LLMs, ensuring high-precision data labeling to reduce algorithmic bias and improve model accuracy.
Streamlined data validation workflows by applying structured evaluation processes to replace manual review, improving overall data processing efficiency.
Contributed to an image domain dataset project, applying structured annotation and quality evaluation workflows to support next-generation computer vision AI systems.

Research Intern – LLM Bias and Benchmarking

TIER Lab, Hunter College – New York, NY

Jun 2025 – Sep 2025

Conducted comparative analysis of 1,000+ prompts across Gemini 2.0 Flash, DeepSeek R1, and GPT-4o to evaluate reasoning consistency, response safety, and model bias.
Designed bias audit protocols for LLM-integrated robotics systems, identifying critical failure modes in Human-Robot Interaction (HRI) safety and fairness.
Engineered reproducible benchmark datasets using Python for robotic perception and dialogue evaluation across multiple model architectures.
Quantified latency and bias trade-offs between high-speed and reasoning-heavy models, delivering data-driven mitigation strategies to research stakeholders.

Capital One First-Gen Focus Fellow

Capital One

Oct 2024 – Dec 2024

Selected for competitive 8-week technical development program; collaborated with software engineers on system design principles and software development lifecycle practices.

QUEST TIMELINE JOURNEY LOG

EXPERIENCE

PROJECTS

EDUCATION

TIP: Select any node to open its full mission log.

EXP

AI Trainer – Software Engineering & Coding Domain

Handshake – Remote

Jun 2025 – Present

Score frontier-LLM responses on multi-turn software-engineering and coding tasks against detailed rubrics, generating high-quality human-feedback data for AI coding assistants.

▸ VIEW DETAILS

EXP

AI Trainer – Image Processing Domain

Handshake – Remote

Nov 2025 – May 2026

Utilized Python-based tooling to evaluate and validate training datasets for LLMs. Streamlined data validation workflows and contributed to image domain dataset projects.

▸ VIEW DETAILS

PROJECT

Onward – AI Emotional-Companion Platform

Next.js 15, TypeScript, Supabase/pgvector, Cerebras GPT-OSS 120B, Gemini

May 2026 – Present

Built a six-lane "FacetsRAG" retrieval pipeline (~95–97% top-1) for emotional-companion matching, deployed on Vercel.

▸ VIEW DETAILS

PAPER

When Directional Accuracy Lies

First-author paper · LoRA-adapted TimesFM · Equity Forecasting

Mar 2026 – Jul 2026 · Published (arXiv)

A base-rate-honest benchmark showing that strong directional accuracy can hide economically worthless equity predictions.

▸ VIEW DETAILS

PROJECT

Weather Forecast App

React, Node.js, Express.js, Vite, REST API

Jul 2025 – Aug 2025

Full-stack weather app with React frontend and Node.js/Express backend consuming OpenWeatherMap API for real-time forecasts.

▸ VIEW DETAILS

EXP

Research Intern – LLM Bias and Benchmarking

TIER Lab, Hunter College

Jun 2025 – Sep 2025

Comparative analysis of 1,000+ prompts across Gemini 2.0 Flash, DeepSeek R1, and GPT-4o. Designed bias audit protocols for LLM-integrated robotics.

▸ VIEW DETAILS

PROJECT

CppAutoSim – Autonomous Driving Simulator

C++, OOP, Algorithm Design

Feb 2025 – Mar 2025

Vehicle simulation engine with realistic kinematics, lane-change algorithms, and obstacle avoidance. Modular OOP architecture.

▸ VIEW DETAILS

EXP

Capital One First-Gen Focus Fellow

Capital One

Oct 2024 – Dec 2024

Competitive 8-week technical development program. Collaborated with engineers on system design and SDLC practices.

▸ VIEW DETAILS

EDU

Hunter College, CUNY

B.A. – Double Major: Computer Science & Mathematics

Expected May 2027

Coursework: Data Structures, Algorithms, Software Analysis & Design, Numerical Methods, Linear Algebra, Discrete Math

▸ VIEW DETAILS

PROJECT ARCHIVE 3 COMPLETE // 1 ACTIVE

PUBLISHED PAPER

When Directional Accuracy Lies

Mar 2026 – Jul 2026 · Published (arXiv)

First-author research introducing a "base-rate-honest" benchmark for LoRA-adapted TimesFM on equity forecasting. Shows that strong directional-accuracy scores can mask economically worthless predictions once base rates are accounted for, and proposes an honest evaluation framework for foundation-model time-series forecasters. Co-authored with SA Kwon.

Python PyTorch TimesFM LoRA Time-Series Financial ML

▶ VIEW REPO // github

Onward – AI Emotional-Companion Platform

May 2026 – Present

Built and deployed a full-stack Next.js 15 app on Vercel with a custom six-lane retrieval-augmented-generation pipeline ("FacetsRAG") combining vector search, reciprocal-rank fusion, and LLM reranking — reaching ~95–97% top-1 accuracy on a hand-graded eval set. Powered by Cerebras GPT-OSS 120B and Gemini embeddings over a pgvector store, with anonymous-first Supabase Auth and per-user/per-IP rate limiting.

Next.js 15 TypeScript Supabase/pgvector Cerebras GPT-OSS 120B Gemini Vercel

▶ LAUNCH APP // onwardapp.me

Weather Forecast App

Jul 2025 – Aug 2025

Full-stack weather app with React frontend and Node.js/Express backend, consuming OpenWeatherMap API for real-time conditions and 5-day forecasts. Features React Context, React Router, Axios, and secure API key management with dotenv.

React Node.js Express.js Vite REST API

CppAutoSim

Feb 2025 – Mar 2025

Vehicle simulation engine in C++ with realistic kinematics including acceleration modeling, lane-change algorithms, and obstacle avoidance logic. Modular OOP codebase supporting scalable integration of new sensor types and environment variables.

C++ OOP Algorithm Design

▶ MORE ON GITHUB

TECH STACK SKILL INVENTORY